EngineX-Ascend/enginex-ascend-910-llama.cpp
10
0
Fork 0
Code Issues Pull Requests Actions 4 Projects Releases Wiki Activity

vulkan: workaround for AMD Windows driver 16 bit unpack8 bug (#12472)

Browse Source
This commit is contained in:
Eve
2025-03-21 19:27:47 +00:00
committed by GitHub
parent af04481e6b
commit 30c42ef5cb
4 changed files with 12 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq3_s.comp
View File

@@ -21,7 +21,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint ib32,
sum[j] = 0.0;
}
[[unroll]] for (uint l = 0; l < 4; ++l) {
const u8vec2 qs = unpack8(data_a_packed16[ibi].qs[4 * ib32 + l]);
const u8vec2 qs = unpack8(uint32_t(data_a_packed16[ibi].qs[4 * ib32 + l])).xy; // vec4 used due to #12147
const uint sign = data_a[ibi].signs[4 * ib32 + l];
const vec4 grid0 = vec4(unpack8(iq3s_grid[qs.x | ((qh << (8 - 2*l)) & 0x100)]));
const vec4 grid1 = vec4(unpack8(iq3s_grid[qs.y | ((qh << (7 - 2*l)) & 0x100)]));